Together AI has disclosed prompt leakage from stored samples of customer requests after a credential for a private repository appeared in a public GitHub project, in a blog post published on 21 November 2026 and signed by chief executive Vipul Ved Prakash. OpenRouter and Novita issued their own notices on 22 November. The three posts follow a 19 November canary-token study by HD Moore, Dragos’s chief research officer. They are separate accounts, and they should be read as such.
What Dragos found with canary tokens
Dragos says it submitted “approximately two million prompts” containing unique URLs and canary-token credentials to widely used AI services between July and November 2026, then watched where those secrets surfaced. It says “data from at least three AI providers was exposed.” Multiple resellers operating through an AI inference marketplace, it adds, “were able to view and retain customer prompts using standard logging capabilities.” It also describes cases of prompts submitted to U.S.-based providers “being transmitted to Chinese AI organizations.” That study is the background to this week’s disclosures, not a substitute for each company’s own notice.
Together AI: prompt leakage from a public credential
Together says that, as part of routine software testing, it collected prompt samples that were authorised for that use and stored them in a private Hugging Face repository. The samples “came from requests to the MiniMax M3, GLM-5.2, and Kimi K2.6 models, primarily on June 24, July 4, July 8, and July 14, 2026.” During development of its public GitHub repository, xorl, “an AI coding agent committed a credential that provided access to the private Hugging Face repository.” Together says automated scanning detected the exposed credential and that it reverted the commit, but that in the exposure window “an unauthorized party downloaded stored prompt samples after a credential for a private repository was exposed in a public code repository.”
The party, Together says, “searched those samples for credentials to other services and tested whether those credentials were valid. Five of the credentials tested were canary credentials associated with a Dragos security experiment.” It has worked “with an external forensic investigation firm and law enforcement.” It says it has not identified evidence of additional prompt leakage beyond this incident, while a review of the affected prompt contents is still under way to determine whether sensitive personal information or other customer data was exposed, and which customers need direct notice.
OpenRouter: no breach of the router, and a provider clean-up
OpenRouter, in a 22 November post, says: “These findings did not show any breach or unauthorized access of OpenRouter.” Its post does not describe prompt leakage on the router itself. It has “Off-boarded three providers from OpenRouter,” and it does not name them. It has also “Removed 125 endpoints from Zero Data Retention (ZDR) eligibility,” including, it says, providers flagged by researchers and further endpoints taken off as a precaution, and has “Notified all potentially affected customers.”
Separately, OpenRouter lists provider announcements it has received: Novita’s MiniMax privacy notice, a preliminary incident report from Phala, and Together’s prompt-leakage post. “Not all affected providers have posted publicly,” it says. That list is not a statement that those three names are the providers it off-boarded. OpenRouter also says it is adapting the researchers’ testing approach so that it can run similar checks on an ongoing basis.
Novita: a routing configuration issue, not a described breach
Novita’s privacy notice describes a different problem. A review, it says, confirmed that “some requests were forwarded to MiniMax’s official API rather than processed on Novita-hosted infrastructure.” The cause, it says, was “A routing configuration issue”: an API forwarding path intended for specific customers applied more broadly. “If you used MiniMax models through Novita between 2026-05-28 and 2026-11-11, your request contents may have been accessed or retained by a third party.” That window includes requests made directly through Novita and those routed to Novita through OpenRouter. Novita says it has suspended the affected models.
It also says “No use of the AWS test key markers was observed.” It is following up with MiniMax about handling of the affected data. That is Novita’s own account of a misdirected routing path. It is not, in the notice, described as a break-in, and nothing in the official posts ties it to the credential exposure Together has set out.
TSN has previously covered Together AI’s post-training tools and OpenRouter’s Auto Code launch, and, on the same credential-hygiene point, GitBook’s advice to rotate API tokens after an AI-agent attack.
Rotate credentials and check ZDR settings
Whether the issue was prompt leakage of stored samples, as Together describes, a routing path that left Novita, or a provider OpenRouter has taken off its roster, the customer action is the same. The practical steps sit in the notices themselves. Anyone who has ever pasted a live credential, API key or other secret into a prompt should rotate it, including for requests sent through a router or a third-party inference host. Together is now making zero data retention the default for new accounts and tells customers handling sensitive prompt data to enable ZDR in their settings. OpenRouter has stripped ZDR eligibility from 125 endpoints, some as a precaution, so customers who relied on a ZDR badge should check whether those routes still qualify. Novita likewise advises revoking or rotating secrets sent in requests. Dragos and OpenRouter both point to ongoing canary-style testing; a clean bill today is not a reason to leave old keys in prompts.
Sources
- Prakash, “Security event: Prompt leakage for requests in June and July of 2026,” Together AI, 21 November 2026.
- OpenRouter, “Data processing incidents on inference providers,” 22 November 2026.
- Novita, “Privacy Notice Regarding MiniMax Requests on Novita,” 22 November 2026.
- Moore, “AI Supply Chain Security: What Canary Tokens Found,” Dragos, 19 November 2026.
Get TSN’s biggest tech stories, explained.
News and insights across AI, tech, crypto and cybersecurity. Every source linked. Unsubscribe anytime.
Get the newsletterRelated stories
- Together AI Lets Teams Write Their Own RL Loops on Open Models is TSN’s earlier account of Together AI’s post-training tools, and is the next read on the company.
- OpenRouter Launches Auto Code and Rebuilds Its Auto Router
- GitBook Users Told to Rotate API Tokens After AI-Agent Attack
- Wiz Counts Nine AI Developer Platform Breaches in ‘Black October’
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
